By name: review description: Review code changes for correctness and quality before merging. Use after implementing a meaningful chunk or feature, between plan tasks, or before opening a PR — to catch issues while they're cheap.
██████╗ ███████╗██╗ ██╗██╗███████╗██╗ ██╗
██╔══██╗██╔════╝██║ ██║██║██╔════╝██║ ██║
██████╔╝█████╗ ██║ ██║██║█████╗ ██║ █╗ ██║
██╔══██╗██╔══╝ ╚██╗ ██╔╝██║██╔══╝ ██║███╗██║
██║ ██║███████╗ ╚████╔╝ ██║███████╗╚███╔███╔╝
╚═╝ ╚═╝╚══════╝ ╚═══╝ ╚═╝╚══════╝ ╚══╝╚══╝
Review on two axes, by a fresh set of eyes
Author-bias elimination
The reviewer must not be the author. A self-review shares the exact blind spots that produced the bug. Dispatch a fresh subagent with crafted context — the diff plus the originating spec/issue — not your session history. The reviewer should reach its verdict without your reasoning leaking in. If you can't spawn a subagent, simulate the reset: review strictly from the diff + spec, deliberately setting aside your implementation reasoning.
Independent axes
Run these reviews so they don't pollute each other (ideally as parallel subagents that never see each other's findings), then present them side by side without merging or reranking:
- Spec — does the change do what was actually asked? Right behavior, all the acceptance criteria, nothing missing or scope-crept.
- Correctness — is the code actually right, independent of the spec? Edge cases, boundary/off-by-one, error and failure paths, null/empty, concurrency and races, resource leaks, injection. (Code can satisfy the spec and follow every convention and still be wrong — this axis is the "correctness" the skill's name promises.)
- Standards — does it follow this repo's conventions and quality bar? Naming, structure, tests, idioms.
Keep them separate because a change routinely passes one and fails another. A merged review lets one mask the others.
First run the repo's own lint / typecheck / test commands and confirm they're green; then the Standards reviewer can skip what tooling already enforces (formatting, lint) and spend its attention on judgment calls. Never skip a check you didn't actually run.
Adversarial framing
Assume there are problems and go find them. "I found zero issues" almost always means you weren't looking hard enough. A clean verdict is only valid if you inspected every changed hunk and can name the single riskiest line you considered and why it's actually safe.
Severity tiers → action
- Critical — security hole, data loss, crash, incorrect output, or a regression. Block; fix before anything else.
- Important — missing test for new logic, an unhandled error path, or a performance cliff. Fix before proceeding.
- Minor — style not enforced by tooling, naming. Note them; don't gate.
Receiving the findings (when feedback comes to you)
- Evaluate each point technically. Don't perform agreement ("Great catch!") — just assess and act.
- If a finding is wrong, push back with reasoning rather than complying blindly.
- If asked to "implement this properly/fully," grep for actual usage first — don't build flexibility nobody needs (
forge:principles#2).
Exit
Fix Critical/Important, then forge:verify with evidence, then forge:ship.